This study proposes a physics-guided neural network-based operational transfer path analysis (OTPA) method that addresses the limitations of conventional OTPA: ill-conditioning, crosstalk, and the absence of frequency response function (FRF) informati...
This study proposes a physics-guided neural network-based operational transfer path analysis (OTPA) method that addresses the limitations of conventional OTPA: ill-conditioning, crosstalk, and the absence of frequency response function (FRF) information. First, a neural network model capable of processing complex-valued data is introduced to accurately estimate transmissibility. Furthermore, an advanced network architecture is designed to incorporate prior known physical information of the system. Specifically, modal parameters—such as natural frequencies, damping ratios, and relative mode shapes—extracted through operational modal analysis are embedded into the neural network, enabling the estimation of system FRFs. To this end, a modified OTPA equation is reformulated to incorporate modal parameters, and the network is trained to reflect the system’s physical behavior. This integrated approach allows the proposed method to perform contribution analysis using the derived FRFs, thereby fundamentally addressing both the ill-conditioning and crosstalk problems. The method is validated using data measured from a test bench resembling an actual vehicle. Vibration and noise responses at indicator and receiver positions were measured under various operating conditions and used to train the model. Compared to conventional OTPA, the proposed method more accurately estimated interior noise. In addition, the FRFs of each transfer path were successfully estimated, and the contribution analysis results showed high consistency with the results of component-based TPA, regardless of the ill-conditioning and crosstalk in the indicator signals. This study presents an effective solution to the theoretical and practical limitations of conventional OTPA. Notably, it enables the identification of dominant transfer paths and their FRFs solely from operating data, which can be effectively utilized to reduce noise through system-level modifications.